Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
KRADA: Known-region-aware Domain Alignment for Open-set Domain Adaptation in Semantic Segmentation
Authors: Chenhong Zhou, Feng Liu, Chen Gong, Rongfei Zeng, Tongliang Liu, William Cheung, Bo Han
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of KRADA has been verified on two synthetic tasks and one COVID-19 segmentation task. Experimental results show that KRADA enables CSDAS methods to identify unknown-class regions and achieve a better overall adaptation, verifying its effectiveness and good generalization ability. We present the results of SYNTHIA Cityscapes in Table 1. Ablation studies: 1) We tend to select the minority classes to construct the unknown class in the above experiments in order to retain as many source images as possible. We also provide some qualitative segmentation results in Figure 4. |
| Researcher Affiliation | Academia | Chenhong Zhou EMAIL Department of Computer Science, Hong Kong Baptist University; Feng Liu EMAIL The University of Melbourne; Chen Gong EMAIL The Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology; Rongfei Zeng EMAIL Northeastern University; Tongliang Liu EMAIL Sydney AI Centre, The University of Sydney; William K. Cheung EMAIL Department of Computer Science, Hong Kong Baptist University; Bo Han EMAIL Department of Computer Science, Hong Kong Baptist University |
| Pseudocode | Yes | Algorithm 1 An implementation of KRADA on existing CSDAS methods. Input: source data (XS, Y S), target data XT , initial pseudo-labels ˆY T 0 . Parameter: network parameters: θF , θC, θD, θC , the number of iteration N, learning rate: µ, pseudo-label loss weight: α, initial threshold: δ, tolerable proportion: γ, momentum factor: β. Output: predicted target labels: e Y T . |
| Open Source Code | Yes | Our source code is available at https://github.com/chenhong-zhou/KRADA |
| Open Datasets | Yes | Based on two synthetic-to-real benchmark tasks in CSDAS: SYNTHIA (Ros et al., 2016) Cityscapes (Cordts et al., 2016) and GTA5 (Richter et al., 2016) Cityscapes, we adjust these two tasks to simulate the OSDAS scenario. To construct a COVID-19 task, we exploit the public datasets summarized in (Ma et al., 2020). The source data consists of normal CT scans with lung annotations. Both target data and test data include COVID-19 cases and non-infected CT scans. More specifically, COVID-19 CT scans are publicly available (Jun et al., 2020) (with CC BY-NC-SA license). |
| Dataset Splits | Yes | Cityscapes is a real-world dataset consisting of a training set with 2,957 images and a validation set with 500 images. We divide the Cityscapes training set into training splits of 2,500 images for training and evaluation splits of 457 images for hyperparameter selection. We divide the target data into two parts for training and testing, and each part includes 10 COVID-19 cases and 5 non-infected CT scans. |
| Hardware Specification | Yes | All the models are implemented using Python 3.6 and Pytorch 1.7 on a TITAN Tesla V100 GPU. |
| Software Dependencies | Yes | All the models are implemented using Python 3.6 and Pytorch 1.7 on a TITAN Tesla V100 GPU. |
| Experiment Setup | Yes | Regarding the pseudo-label hyperparameters in KRADA, γ, β, and δ are chosen as 0.05%, 0.99, and 0.1 respectively for these two synthetic segmentation tasks. α is set as 0.1, 0.03, and 0.2 for the Adapt Seg Net, CLAN, and FADA models equipped with KRADA. In this task, the pseudo-label hyperparameters γ, β, and δ are chosen as 1%, 0.99, and 0.001 respectively. α is set as 0.01, 0.01, and 0.1 for the three models equipped with KRADA. |